Identity-consistent palm vein image generation method and system

Abstract

The present application provides a method and system for generating palm vein images with identity consistency, the method comprising the steps of: collecting a number of real palm vein images and performing preprocessing to obtain a number of palm vein sample images; constructing a pseudo palm contour and randomly constructing a number of pseudo vein lines in each pseudo palm contour to obtain a palm vein template image; constructing a palm vein image generation model, and inputting the palm vein sample image and the palm vein template image into the palm vein image generation model for training; inputting the palm vein template image into the trained palm vein image generation model, and generating pseudo palm vein data set samples by randomly introducing different noise vectors z and performing feature fusion with the palm vein template image. The method of the present application can generate a large number of pseudo palm vein data set samples with identity consistency and taking into account diversity.

Description

Technical Field

[0001] The present application belongs to the field of palm vein recognition technology, and more specifically, to a method and system for generating an identity-consistent palm vein image. Background Art

[0002] Palm vein recognition technology uses the uniqueness of the veins inside the palm to identify the person. Palm vein recognition technology mainly consists of image acquisition, image preprocessing, image feature extraction, and feature matching. Palm vein recognition technology uses a non-contact recognition method. The user only needs to place the palm on the recognition device to complete the recognition without direct contact with the device. Although the non-contact collection method avoids the user's concerns about privacy and hygiene, it is very difficult to collect a large amount of palm vein image data in a short time because it requires a high degree of cooperation and is easily affected by environmental interference. For this reason, palm vein data sets are very scarce and rare in public scenarios, and are very expensive to collect, which further limits the development of palm vein recognition technology.

[0003] In the existing technology, the image generation model based on deep learning is mainly based on GAN network, and has been widely used in face recognition scenarios, such as face image generation, face image enhancement and super-resolution, face attribute editing, face image style conversion and other fields. Although the generative adversarial network has shown strong generation capabilities in the field of face recognition, palm vein image data is more sensitive than facial images and contains complex textures and subtle vascular structures. Due to the lack of a large amount of palm vein data in real scenes, it is currently impossible for GAN to learn and accurately reproduce these details, but these details are crucial for identity recognition. This leads to the situation that if the training data is overfitted, the generated samples are limited to specific patterns in the training set, and only palm vein images of a single identity (ID) can be generated, rather than truly generating novel and diverse samples. Diversity refers to posture, tilt angle, different brightness, clarity, etc. If the training data focuses on generalization ability, it is impossible to generate palm vein images of a single identity (ID). Summary of the invention

[0004] The purpose of the embodiments of the present application is to provide a method and system for generating a palm vein image with identity consistency, so as to solve the technical problem that the prior art cannot generate a pseudo palm vein dataset sample with identity consistency and diversity in the palm vein image generation process.

[0005] To achieve the above purpose, the technical solution adopted by the present application is: to provide a method for generating an identity-consistent palm vein image, comprising the steps of:

[0006] Collect a number of real palm vein images and perform preprocessing to obtain a number of palm vein sample images;

[0007] Constructing a pseudo palm contour and randomly constructing a number of pseudo vein lines in each pseudo palm contour to obtain a palm vein template image;

[0008] Constructing a palm vein image generation model, and inputting the palm vein sample image and the palm vein template image into the palm vein image generation model for training;

[0009] The palm vein sample image is downsampled and mapped to the Gaussian noise domain to obtain a large number of high-dimensional sample feature codes, and a feature coding loss is established to constrain the sample feature code to a noise vector z randomly sampled from the Gaussian space domain;

[0010] The palm vein template image is fused with the sample feature code and the noise vector z to generate a pseudo palm vein sample image B' and a pseudo palm vein sample image B", and the cosine similarity between the pseudo palm vein sample image B' and the pseudo palm vein sample image B" is constrained by identity consistency loss;

[0011] The palm vein template image is input into the trained palm vein image generation model, and the pseudo palm vein dataset samples are generated by randomly introducing different noise vectors z and fusing the features with the palm vein template image.

[0012] In another preferred embodiment of the present application, the palm vein image generation model includes:

[0013] Encoder F, used for downsampling the palm vein sample image and mapping it to the Gaussian noise domain to obtain a high-dimensional sample feature encoding;

[0014] A generator G is used to perform feature fusion on the palm vein template image, the sample feature code, and the noise vector z to generate a pseudo palm vein sample image B' and a pseudo palm vein sample image B"; and

[0015] The recognition model D is used to extract the identity characteristics of the pseudo palm vein sample image B' and the pseudo palm vein sample image B".

[0016] In another preferred embodiment of the present application, a method for generating a pseudo palm vein sample image B' by performing feature fusion of a palm vein template image and a sample feature code comprises the steps of:

[0017] The spatial dimension of the palm vein template image is gradually reduced through the generator G during the downsampling process;

[0018] In each spatial dimension, the template image feature map of the palm vein template image and a random sample feature code are fused to obtain a low-dimensional fused feature map;

[0019] The spatial dimension of the fused feature map is increased by upsampling to generate a pseudo palm vein sample image B'.

[0020] In another preferred embodiment of the present application, before constructing the palm vein image generation model, the following steps are further included:

[0021] Construct recognition model D and use arcface as loss function;

[0022] In the training stage, the palm vein sample images are input into the recognition model D as the training set, and the ArcFace loss function is calculated to make the angular distance between feature vectors of different categories larger and the feature vectors of the same category closer;

[0023] In the inference stage, the palm vein sample image is input into the trained recognition model D to obtain the feature vector of the palm vein sample image. The identity consistency is determined by calculating the cosine similarity between the feature vectors of different palm vein sample images.

[0024] In another preferred embodiment of the present application, the calculation formula of the feature coding loss is:

[0025]

[0026] Where L Feature represents the feature coding loss, x represents the high-dimensional sample feature coding output by the encoder F, z represents the noise vector, z~Pgassion(z) represents that the noise vector z follows the Gaussian distribution, x~Pdata(x) represents that the sample feature coding x follows the original data distribution of the palm vein sample image, E z～Pgassion（z），x～Pdata(x) represents the mathematical expectation value based on the joint distribution between the sample feature code x and the noise vector z, and ||zx|| represents the L1 distance between the sample feature code x and the noise vector z.

[0027] In another preferred embodiment of the present application, the calculation formula of the identity consistency loss is:

[0028]

[0029] Where L ID represents identity consistency loss, “·” represents vector dot product operation, “×” represents modular multiplication operation, and D(B’) and D(B”) are the identity feature vectors extracted by the trained recognition model D from the pseudo palm vein sample image B’ and the pseudo palm vein sample image B”, respectively.

[0030] In another preferred embodiment of the present application, after the palm vein template image is feature-fused with the sample feature code and the noise vector z to generate the pseudo palm vein sample image B' and the pseudo palm vein sample image B", the following steps are also included:

[0031] The pseudo palm vein sample image B' generated by the image content loss constraint generator G retains the overall content information in the palm vein sample image.

[0032] In another preferred embodiment of the present application, the calculation formula for the image content loss is:

[0033]

[0034] Where L Content represents the image content loss, B i ' represents the i-th pixel value in the pseudo palm vein sample image B', B i represents the i-th pixel value in the palm vein template image, and N represents the total number of pixels in the pseudo palm vein sample image B' and the palm vein sample image.

[0035] In another preferred embodiment of the present application, after establishing the image content loss, the following steps are further included:

[0036] Calculate the overall loss of the palm vein image generation model. The calculation formula of the overall loss is as follows:

[0037]

[0038] Where L total represents the overall loss, λ Feature , Content and λ ID They are feature encoding loss L Feature , image content loss L Content and identity consistency loss L ID The weight value of .

[0039] The present application also provides a palm vein image generation system, the palm vein image generation system comprising:

[0040] A template construction unit, used to construct a pseudo palm contour and randomly construct a number of pseudo vein lines in each pseudo palm contour to obtain a palm vein template image; and

[0041] The image generation unit is used to input the palm vein template image into the palm vein image generation model as described above to generate a pseudo palm vein data set sample.

[0042] Compared with the prior art, the identity-consistent palm vein image generation method provided in the present application obtains a palm vein template image by constructing a pseudo palm contour and randomly constructing a number of pseudo vein lines in each pseudo palm contour, and at the same time establishes feature encoding loss and identity consistency loss to train the palm vein image generation model, and generates a large number of pseudo palm vein data set samples with identity consistency and diversity by inputting a noise vector z.

[0043] The palm vein image generation system provided in the present application, by adopting the palm vein image generation model as described above, can automatically generate a large number of pseudo palm vein dataset samples with identity consistency and diversity, thereby facilitating the expansion of dataset samples in the field of palm vein recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0045] Figure 1 A schematic diagram of the process of generating a palm vein image with identity consistency provided in an embodiment of the present application;

[0046] Figure 2 A schematic diagram of a palm vein sample image provided in an embodiment of the present application;

[0047] Figure 3 A schematic diagram of a palm vein template image provided in an embodiment of the present application;

[0048] Figure 4 A schematic diagram of a pseudo palm vein dataset sample provided in an embodiment of the present application;

[0049] Figure 5 A structural diagram of a palm vein image generation model provided in an embodiment of the present application;

[0050] Figure 6 A schematic diagram of the reasoning process of the recognition model D provided in the embodiment of the present application;

[0051] Figure 7 A schematic diagram of feature fusion of a palm vein template image and a sample feature code provided in an embodiment of the present application. DETAILED DESCRIPTION

[0052] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0053] It should be noted that when an element is referred to as being "fixed to" or "disposed on" another element, it can be directly on the other element or indirectly on the other element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or indirectly connected to the other element.

[0054] It should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0055] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0056] Please also read Figures 1 to 4 The method for generating a palm vein image with identity consistency provided in an embodiment of the present application is now described. The method for generating a palm vein image with identity consistency comprises:

[0057] Image acquisition step S1: acquiring a number of real palm vein images and performing preprocessing to obtain a number of palm vein sample images.

[0058] It can be understood that in the image acquisition step S1, the real palm vein images used can come from the palms of multiple people, and each palm is based on different postures and angles. In this way, multiple real palm vein images from the same palm have both identity consistency and diversity.

[0059] The preprocessing may include but is not limited to palm vein ROI extraction, normalization, denoising, image enhancement, rotation and cropping, etc. These steps act together on the real palm vein image, which is intended to improve the image quality and facilitate the subsequent sample feature coding extraction.

[0060] Template construction step S2: construct a pseudo palm contour and randomly construct a number of pseudo vein lines in each pseudo palm contour to obtain a palm vein template image.

[0061] The template construction step S2 of this application is based on the analysis of the palm vein vascular structure of a large number of individuals, and it is determined that the palm vein vascular structure mainly presents a random divergent distribution pattern, which is composed of many small vascular branches, that is, the location of the blood vessels is random, and the blood vessels are interconnected and divergent. These blood vessels eventually converge into several main veins, and finally present a network structure as a whole. Usually, the palm vein patterns of different individuals are similar to lines of different lengths and curvatures, and the number of lines generally ranges from 1 to 20.

[0062] Therefore, in the template construction step S2, the pseudo palm contour can be depicted according to the palm vein sample image, and a number of pseudo vein lines can be automatically generated and randomly constructed in each pseudo palm contour by manual or AI. The pseudo vein lines match the palm vein vascular structure. The pseudo vein line features in each palm vein template image obtained in this way are unique, so each palm vein template image has a unique pseudo identity feature.

[0063] Model training step S3.1: constructing a palm vein image generation model, and inputting the palm vein sample image and the palm vein template image into the palm vein image generation model for training;

[0064] Model training step S3.2: downsample the palm vein sample image and map it to the Gaussian noise domain to obtain a large number of high-dimensional sample feature codes, and establish a feature coding loss to constrain the sample feature code to the noise vector z randomly sampled from the Gaussian space domain;

[0065] Model training step S3.3: perform feature fusion of the palm vein template image with the sample feature code and the noise vector z to generate a pseudo palm vein sample image B' and a pseudo palm vein sample image B", and establish a cosine similarity between the pseudo palm vein sample image B' and the pseudo palm vein sample image B" constrained by identity consistency loss;

[0066] The ultimate goal of this application is to obtain a large number of pseudo palm vein dataset samples that have identity consistency and diversity.

[0067] Therefore, in the model training steps S3.1-S3.3, a palm vein image generation model is first constructed, and the palm vein sample image is downsampled and mapped to the Gaussian noise domain through the palm vein image generation model to obtain a large number of high-dimensional sample feature codes. The sample feature codes obtained in this way can be aligned with the noise codes of the noise vector z. Furthermore, in the training process, because the sample feature codes are constrained to the noise vector z randomly sampled from the Gaussian space domain based on the feature coding loss, the difference between the noise code of the noise vector z and the sample image code can be gradually reduced by relying on the feature coding loss. When the feature coding loss converges, the noise coding can be controlled to generate the desired pseudo palm vein sample image.

[0068] At the same time, under the constraint of feature coding loss, since the data distribution of sample feature coding gradually approaches the noise vector z, when the palm vein template image and the noise vector z are feature fused, a pseudo palm vein sample image B” that is gradually similar to the pseudo palm vein sample image B’ can be generated, thereby gradually increasing the ability to generate diversity during the training process. On this basis, since the identity consistency loss can constrain the cosine similarity between the pseudo palm vein sample image B’ and the pseudo palm vein sample image B”, that is, during the training process, the identity consistency loss is used to provide identity consistency constraints for the pseudo palm vein sample image B’ and the pseudo palm vein sample image B”.

[0069] Therefore, the feature encoding loss and the identity consistency loss work together to increase the randomness and diversity of the generated pseudo palm vein sample images, and generate adversarial protection for identity consistency.

[0070] Image generation step S4: input the palm vein template image into the trained palm vein image generation model, and generate pseudo palm vein data set samples by randomly introducing different noise vectors z and fusing the features with the palm vein template image.

[0071] In the image generation step S4, after each palm vein template image is input into the trained palm vein image generation model, a pseudo palm vein data set sample with identity consistency and diversity can be generated. Based on the template construction step S2, on the premise of unlimited generation of palm vein template images, a large number of pseudo palm vein data set samples with identity consistency and diversity can be expanded in the field of palm vein recognition, overcoming the problem of lack of a large number of public data sets in the field of palm vein recognition technology.

[0072] In another embodiment of the present application, please refer to Figure 5 , the palm vein image generation model includes:

[0073] Encoder F, used for downsampling the palm vein sample image and mapping it to the Gaussian noise domain to obtain a high-dimensional sample feature encoding;

[0074] A generator G is used to perform feature fusion on the palm vein template image, the sample feature code, and the noise vector z to generate a pseudo palm vein sample image B' and a pseudo palm vein sample image B"; and

[0075] The recognition model D is used to extract the identity characteristics of the pseudo palm vein sample image B' and the pseudo palm vein sample image B".

[0076] It can be understood that, for example, the encoder F adopts the resnet18 network structure, and the palm vein sample image can be downsampled multiple times and mapped to the Gaussian domain through the encoder F, so that the sample feature coding is aligned with the noise coding of the noise vector z, thereby improving the accuracy of calculating the feature coding loss.

[0077] The generator G can adopt an encoding-decoding or downsampling-upsampling structure. The encoding part uses convolutional layers and pooling layers to capture the identity features of the palm vein template image and gradually reduce the spatial dimension, while the decoding part increases the spatial dimension through a deconvolution layer (or upsampling layer) to randomly generate realistic pseudo palm vein sample images.

[0078] The recognition model D is a pre-trained recognition model, so that after the pseudo palm vein sample image B′ and the pseudo palm vein sample image B″ are input into the recognition model D, the identity characteristics can be extracted.

[0079] For further information, please also refer to Figure 6 The method of generating a pseudo palm vein sample image B' by fusing the palm vein template image with the sample feature code comprises the steps of:

[0080] The spatial dimension of the palm vein template image is gradually reduced through the generator G during the downsampling process;

[0081] In each spatial dimension, the template image feature map of the palm vein template image and a random sample feature code are fused to obtain a low-dimensional fused feature map;

[0082] The spatial dimension of the fused feature map is increased by upsampling to generate a pseudo palm vein sample image B'.

[0083] It can be understood that, unlike the function of the generator G of the traditional U-Net network structure, the embodiment of the present application first extracts the high-level, abstract features in the palm vein template image through downsampling during the downsampling process. These features usually contain key information that can represent the identity in the palm vein template image, such as vein texture, shape contour, etc. Since this feature fusion is performed dimension by dimension, the generated low-dimensional fusion feature map not only contains the diversity information of the random sample, but also retains the key features of the palm vein template image. Although random sample feature encoding is introduced in the generation process, its main function is to provide diversity and randomness without changing or destroying the template image feature map. Through feature fusion, the random sample feature encoding interacts with the template image feature map and jointly affects the generation result, but the features of the template image always dominate the generation process.

[0084] For further information, please also refer to Figure 7 In the model training step S3.1, before constructing the palm vein image generation model, the following steps are also included:

[0085] Construct recognition model D and use arcface as loss function;

[0086] In the training stage, the palm vein sample images are input into the recognition model D as the training set, and the ArcFace loss function is calculated to make the angular distance between feature vectors of different categories larger and the feature vectors of the same category closer;

[0087] In the inference stage, the palm vein sample image is input into the trained recognition model D to obtain the feature vector of the palm vein sample image. The identity consistency is determined by calculating the cosine similarity between the feature vectors of different palm vein sample images.

[0088] It is understandable that for the palm vein image generation method with identity consistency proposed in the present invention, the generated pseudo palm vein data set samples should have identity consistency while ensuring diversity. Therefore, before training the palm vein image generation model, the present invention needs to pre-train the recognition model D to provide label information of identity features for the subsequently generated pseudo palm vein sample images. In this way, during the training stage of the palm vein image generation model, the greater the cosine similarity, the greater the possibility that the two subsequently generated pseudo palm vein sample images belong to the same identity feature, and vice versa. The greater the possibility that they belong to different identity features. By obtaining the identity label of the pseudo palm vein sample image through the pre-trained recognition model D, the pseudo palm vein sample image generated by the subsequent palm vein image generation model has randomness and diversity, and will not destroy the intra-class identity consistency.

[0089] Furthermore, in the model training step S3.2, the calculation formula of the feature encoding loss is:

[0090]

[0091] Where L Feature represents the feature coding loss, x represents the high-dimensional sample feature coding output by the encoder F, z represents the noise vector, z~Pgassion(z) represents that the noise vector z follows the Gaussian distribution, x~Pdata(x) represents that the sample feature coding x follows the original data distribution of the palm vein sample image, E z～Pgassion（z），x～Pdata(x) represents the mathematical expectation value based on the joint distribution between the sample feature code x and the noise vector z, and ||zx|| represents the L1 distance between the sample feature code x and the noise vector z.

[0092] It can be understood that the feature encoding loss L in this embodiment is Feature The goal is to minimize the data distribution difference between the sample feature encoding output by the encoder F and the noise vector z, that is, the feature encoding loss constrains the sample feature encoding to the noise vector z randomly sampled from the Gaussian space domain so that the target domain of the encoder F remains approximately normal distribution.

[0093] Furthermore, in the model training step S3.3, the calculation formula of the identity consistency loss is:

[0094]

[0095] Where L ID represents identity consistency loss, “·” represents vector dot product operation, “×” represents modular multiplication operation, and D(B’) and D(B”) are the identity feature vectors extracted by the trained recognition model D from the pseudo palm vein sample image B’ and the pseudo palm vein sample image B”, respectively.

[0096] It can be understood that by calculating the cosine similarity of the identity feature vectors between the pseudo palm vein sample image B' and the pseudo palm vein sample image B" through the identity consistency loss, the similarity between the generated pseudo palm vein sample image B' and the pseudo palm vein sample image B" can be constrained and shortened. Therefore, the existence of the identity consistency loss ensures the randomness and diversity of the palm veins generated by the generator G without destroying the intra-class identity consistency.

[0097] For further information, please also refer to Figure 5 In the model training step S3.3, after the palm vein template image is feature-fused with the sample feature code and the noise vector z to generate the pseudo palm vein sample image B' and the pseudo palm vein sample image B", the following steps are also included:

[0098] The pseudo palm vein sample image B' generated by the image content loss constraint generator G retains the overall content information in the palm vein sample image.

[0099] Furthermore, the calculation formula of the image content loss is:

[0100]

[0101] Where L Content represents the image content loss, B i ' represents the i-th pixel value in the pseudo palm vein sample image B', B i represents the i-th pixel value in the palm vein template image, and N represents the total number of pixels in the pseudo palm vein sample image B' and the palm vein sample image.

[0102] It can be understood that the overall content information such as color, shape, brightness, etc., the generator G is guided to generate an image that is closer to the distribution of the palm vein sample image, thereby improving the authenticity of the image generated by the generator G.

[0103] Furthermore, after establishing the image content loss, the following steps are also included:

[0104] Calculate the overall loss of the palm vein image generation model. The calculation formula of the overall loss is as follows:

[0105]

[0106] Where L total represents the overall loss, λ Feature , Content and λ ID They are feature encoding loss L Feature , image content loss L Content and identity consistency loss L ID The weight value of .

[0107] It is understandable that by establishing the overall loss L total During the training process, the weight value λ can be dynamically adjusted according to the loss change Feature , Content and λ ID This ensures that the palm vein image generation model can take into account the consistency, diversity and authenticity of the ID card during the training process.

[0108] The present application also provides a palm vein image generation system, the palm vein image generation system comprising:

[0109] A template construction unit, used to construct a pseudo palm contour and randomly construct a number of pseudo vein lines in each pseudo palm contour to obtain a palm vein template image; and

[0110] The image generation unit is used to input the palm vein template image into the palm vein image generation model as described above to generate a pseudo palm vein data set sample.

[0111] The palm vein image generation system provided in the present application, by adopting the palm vein image generation model as described above, can automatically generate a large number of pseudo palm vein dataset samples with identity consistency and diversity, thereby facilitating the expansion of dataset samples in the field of palm vein recognition.

[0112] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method for generating an identity-consistent palm vein image, characterized in that: Includes steps: Collect a number of real palm vein images and perform preprocessing to obtain a number of palm vein sample images; Constructing a pseudo palm contour and randomly constructing a number of pseudo vein lines in each pseudo palm contour to obtain a palm vein template image; Constructing a palm vein image generation model, and inputting the palm vein sample image and the palm vein template image into the palm vein image generation model for training; The palm vein sample image is downsampled and mapped to the Gaussian noise domain to obtain a large number of high-dimensional sample feature codes, and a feature coding loss is established to constrain the sample feature code to a noise vector z randomly sampled from the Gaussian space domain; The palm vein template image is fused with the sample feature code and the noise vector z to generate a pseudo palm vein sample image B' and a pseudo palm vein sample image B", and the cosine similarity between the pseudo palm vein sample image B' and the pseudo palm vein sample image B" is constrained by identity consistency loss; The palm vein template image is input into the trained palm vein image generation model, and the pseudo palm vein dataset samples are generated by randomly introducing different noise vectors z and fusing the features with the palm vein template image. The palm vein image generation model includes: Encoder F, used for downsampling the palm vein sample image and mapping it to the Gaussian noise domain to obtain a high-dimensional sample feature encoding; A generator G is used to perform feature fusion on the palm vein template image, the sample feature code, and the noise vector z to generate a pseudo palm vein sample image B' and a pseudo palm vein sample image B"; and Recognition model D, used to extract identity characteristics of pseudo palm vein sample image B' and pseudo palm vein sample image B"; The method for generating a pseudo palm vein sample image B' by fusing a palm vein template image with a sample feature code comprises the following steps: The spatial dimension of the palm vein template image is gradually reduced through the generator G during the downsampling process; In each spatial dimension, the template image feature map of the palm vein template image and a random sample feature code are fused to obtain a low-dimensional fused feature map; The spatial dimension of the fused feature map is increased by upsampling to generate a pseudo palm vein sample image B'.

2. The method for generating an identity-consistent palm vein image according to claim 1, characterized in that: Before constructing the palm vein image generation model, the following steps are also included: Construct recognition model D and use arcface as loss function; In the training stage, the palm vein sample images are input into the recognition model D as the training set, and the ArcFace loss function is calculated to make the cosine distance between feature vectors of different categories larger and the feature vectors of the same category closer; In the inference stage, the palm vein sample image is input into the trained recognition model D to obtain the feature vector of the palm vein sample image. The identity consistency is determined by calculating the cosine similarity between the feature vectors of different palm vein sample images.

3. The method for generating an identity-consistent palm vein image according to claim 1, characterized in that: The calculation formula of the feature encoding loss is: ， Where L Feature represents the feature coding loss, x represents the high-dimensional sample feature coding output by the encoder F, z represents the noise vector, z~Pgassion(z) represents that the noise vector z follows the Gaussian distribution, x~Pdata(x) represents that the sample feature coding x follows the original data distribution of the palm vein sample image, E z～Pgassion（z），x～Pdata(x) represents the mathematical expectation value based on the joint distribution between the sample feature code x and the noise vector z, and ||zx|| represents the L1 distance between the sample feature code x and the noise vector z.

4. The method for generating an identity-consistent palm vein image according to claim 3, characterized in that: The calculation formula of the identity consistency loss is: ， Where L ID represents identity consistency loss, "·" represents vector dot product operation, "×" represents modular multiplication operation, D(B') and D(B") are the identity feature vectors extracted by the trained recognition model D from the pseudo palm vein sample image B' and the pseudo palm vein sample image B", respectively.

5. The method for generating an identity-consistent palm vein image according to claim 4, characterized in that: After the palm vein template image is subjected to feature fusion with the sample feature code and the noise vector z to generate the pseudo palm vein sample image B' and the pseudo palm vein sample image B", the method further includes the following steps: The pseudo palm vein sample image B' generated by the image content loss constraint generator G retains the overall content information in the palm vein sample image.

6. The method for generating an identity-consistent palm vein image according to claim 5, characterized in that: The calculation formula of the image content loss is: ， Where L Content represents the image content loss, B i ' represents the i-th pixel value in the pseudo palm vein sample image B', B i represents the i-th pixel value in the palm vein template image, and N represents the total number of pixels in the pseudo palm vein sample image B' and the palm vein sample image.

7. The method for generating an identity-consistent palm vein image according to claim 6, wherein: After establishing the image content loss, the following steps are also included: Calculate the overall loss of the palm vein image generation model. The calculation formula of the overall loss is as follows: ， Where L total represents the overall loss, λ Feature , Content and λ ID They are the feature encoding loss L Feature , image content loss L Content and identity consistency loss L ID The weight value of .

8. A palm vein image generation system, characterized in that: The palm vein image generation system comprises: A template construction unit, used to construct a pseudo palm contour and randomly construct a number of pseudo vein lines in each pseudo palm contour to obtain a palm vein template image; and An image generating unit is used to input a palm vein template image into a palm vein image generating model in the identity consistent palm vein image generating method as claimed in any one of claims 1 to 7 to generate a pseudo palm vein data set sample.

Images